A marketing team runs K-means on customers described by three behavioural features — spend, purchase frequency and recency. The coloured spheres are those customers, positioned at their true feature values and coloured by their K-means segment. This lab shows the same clustered data through two different lenses used to visualise it.
Because t-SNE optimises local neighbour probabilities rather than preserving a fixed coordinate system, the axes of a t-SNE plot carry no independent meaning and cluster sizes/distances shouldn't be compared directly — only cluster membership is reliable. PCA axes, by contrast, are physically interpretable directions of maximum variance.
The same K-means-clustered customer cloud, shown two ways: a PCA rotation that keeps every true distance intact, and a t-SNE-style layout that exaggerates local neighbourhoods while discarding global distance.
PCA rotates the data onto the axes of maximum variance — a rigid transform, so distances between any two customers (and segments) are preserved exactly. t-SNE instead pulls each segment toward its own local neighbourhood and pushes segments apart, so inter-segment distances become meaningless even though membership stays reliable.
Drag the View slider to morph between the PCA and t-SNE layouts. Change the number of segments, the t-SNE perplexity, and how separated the underlying customer behaviours really are, and watch the two projections respond differently.
Because t-SNE's axes carry no fixed meaning, cluster sizes and the gaps between clusters in a t-SNE plot shouldn't be read literally — only which points end up grouped together is trustworthy, unlike PCA's physically interpretable variance axes.